Work Package 1: Project management
This work package covered the scientific and administrative coordination of the project. It included planning and monitoring of research activities, supervision and coordination with host institutions, reporting to the funding authority, and compliance with ethical and open science requirements. Regular meetings and administrative support ensured that the project progressed according to plan and adapted smoothly to changes in research focus.
Work Package 2: Design an Autotelic DreamCoder
This work package investigated mechanisms for autotelic learning, focusing on systems that can generate, select, and pursue their own goals without predefined task sequences. Rather than developing a single system, the work explored multiple approaches to goal generation, intrinsic motivation, and program-based representations. A key result was ACES, which demonstrated how program representations can be used to generate diverse and progressively challenging programming tasks by targeting skill descriptions and calibrating difficulty based on solver performance.
Work Package 3: Design an Interactive Autotelic DreamCoder
This work package studied how interaction and language shape learning. Computational models were developed in which artificial agents combine trial-and-error experience with natural-language guidance from humans or other agents. The work showed how advice can guide exploration, shape learned world models, and influence learning performance. Methods were also introduced in which agents express successful strategies as short linguistic rules, supporting interpretability and coordination.
Work Packages 4–6: Human experiments and analysis
These work packages designed, ran, and analyzed online experiments with human participants. The experiments examined how human input influences exploration strategies, learning dynamics, and world modeling in artificial systems. The results showed that human guidance can significantly shape learning behavior and, in some settings, that knowledge transfer can occur in both directions, with artificial agents also producing guidance usable by humans.
Work Package 7: Design a real-world use case
The real-world application initially planned in the project was not implemented. During the project, the focus shifted toward developing general learning mechanisms and controlled experimental studies rather than a single end-user application.
Work Package 8: Training, dissemination, and communication
Throughout the project, results were disseminated through peer-reviewed publications, conference presentations, workshops, and open-source releases. The project also contributed to training through supervision of students and interdisciplinary collaboration across machine learning, cognitive science, and social sciences.
